Prompts Are Maps, Not the Territory: Getting the AI Result You Actually Want
Vague prompts and bloated contexts cost real time and money: repeated iterations, misinterpreted requirements, and outputs that miss the actual objective. These essays turn that failure into a checklist: find your unknowns, prune your system prompts, and check whether the goal itself is too small.
Three recent essays from Anthropic engineers Thariq Shihipar and Felix Rieseberg pin most AI disappointments on the requester, not the model. Shihipar compares prompts and background info to a map: the real work is the terrain, and unmarked territory forces AI to guess. He breaks missing knowledge into four categories, from known knowns to unknown unknowns, and pushes for blind-spot checks, interviews, and prototypes to surface them. For photo editing, that means asking AI to explain problems and produce several distinct versions instead of repeating "make it more sophisticated."
Shihipar's second essay takes aim at context bloat. Anthropic cut more than 80% of Claude Code's system prompts for newer models and measured no loss in coding evals, because old constraints pile up and conflict. The advice: keep context lightweight and validate any deletion by running comparison tests on representative tasks. The 80% number only holds for their model and eval setup, so copying it blindly would be a mistake.
Rieseberg's essay pushes the goal itself. He splits ambition into speed, height, depth, and breadth. Speed is just doing the same thing faster. Height means going from "make a ten-page PPT" to "what decision should the meeting reach." Depth means professional method: counterexamples, viewpoint shifts, questions that require real understanding. Breadth means touching tools and fields you'd normally skip. Professional tone, he warns, is not professional method; it needs data, tools, and verification.
The "delete 80% of system prompts" experiment reframes prompt bloat as a liability rather than a safeguard, while also warning that copying the number without matching model and eval context is meaningless.
The four-category unknown framework explains why prototyping beats specification for tacit preferences: you often recognize what you want only after seeing a concrete version.
The distinction between smooth conversation and correct answers, drawn from the photo-editing example, undermines the habit of treating AI chat as authoritative.
The ambition axes imply that most AI use today stays on the speed axis, leaving height, depth, and breadth untapped because they require extra steps and verification.